HadGEM3

OASIS HadGEM-G31 coupled model benchmark

The models UM and Nemo are coupled. They have to exchange fields every one model-hour, so if one model runs an hour faster than the other it will need to wait for the other model to catch up, so they can exchange fields and carry on. If you speed up the slower model the coupled job will run faster, but if you up the faster model there won't be any change in total run time.

Build Instructions

The build process followed on each of the systems benchmarked (if available):

• ARCHER(EPCC)

Example job scripts

Example job scripts can be found here.

Results

Benchmark results

• ARCHER

The output

Two of the output file on ARCHER. To check if the run was successfull, see coupled_run.oxxx:

• Usually the model will either crash or run with NaNs if something goes wrong.
• In the UM output check that the following lines don't vary too much:

This timestep                         This run
Min theta1     proc          position            Min theta1 timestep
224.76    1712   114.6deg W      83.0deg N       223.38   132
Largest negative delta theta1 at minimum theta1
This timestep =    -0.07K. At min for run =    -3.60K

Maximum vertical velocity at timestep  432       Max w this run
w_max   level  proc         position             run w_max level timestep
0.223E+01  42    747  149.8deg W    -14.9deg S    0.753E+01   40   140

• For the following lines the number of iterations shouldn't get too large:

********************************************
*    Linear solve for Helmholtz problem    *
* Outer Inner Iterations InitialError      *
*    2     1       12     0.197758E+00     *
*    2     2        3     0.151540E-02     *
********************************************

• Statistics re the job can be found in pe_output/atmos.fort6.pe0 as well as the UM Run job-start and end time.
• There are many more other files generated during the run.